AI Infrastructure: Gartner Adds OVHcloud and Scaleway to Its Radar

Seasonal news: there is now a Magic Quadrant for “AI cloud infrastructures.”

The outcome was predictable. The latest Magic Quadrant for public cloud infrastructure (IaaS + PaaS), published in summer 2025, had already set the tone. Gartner highlighted the AI component at length, from Microsoft’s reliance on OpenAI to AWS’s lack of competitiveness in generative models.

In parallel, the Magic Quadrant for “AI cloud services for developers” gave way to the broader one for “AI application development platforms.” The majority of vendors ranked in one were found in the other.

OVHcloud as a “Challenger”, Scaleway as a “Niche Player”

One of the rare moments in a single Magic Quadrant where two French providers appear.

OVHcloud is sufficiently advanced on the so‑called execution axis (which reflects the ability to respond to demand) to be positioned among the “Challengers.” Gartner notes a singular advantage in the competitive landscape: the absence of exit fees. It also lauds OVHcloud’s control over its server manufacturing and, more broadly, its now-profitable business.

OVHcloud does not boast the advanced networking capabilities of its main rivals. This limits options for distributed workloads. It also lacks operational-management features, notably for cost governance. It does not have the same catalog of serverless options, MLOps services, and PaaS integrations. This could imply additional integration efforts.

Compared with the other vendors classified as “niche players,” Scaleway offers a broader portfolio (VMs, bare metal, PaaS, serverless…). Gartner sees this as a way to minimize latency and data-transfer costs between “AI components and non‑AI components.” It also praises energy-efficient data centers. And the availability of managed services (Kapsule, Managed Inference…) that reduce the need for MLOps skills.

Scaleway’s cloud regions are exclusively in Europe: difficult to rely on for global workloads that require very low latency. As with OVHcloud, the native PaaS range and integration catalog are less comprehensive than those of the major hyperscalers. Moreover, while the base pricing is “transparent,” the separate billing for storage, instances, and bandwidth can complicate cost management.

17 providers, 6 “Leaders”

Gartner identifies six “leaders” in this segment of AI cloud infrastructures: Alibaba Cloud, AWS, Google, Huawei Cloud, Microsoft and Oracle.

On the execution axis, the situation is as follows:

On the “vision” axis, which translates strategies (commercial, marketing, innovation…):

Interface et documentation complexes chez Alibaba Cloud

Alibaba Cloud stands out for the comprehensiveness of its portfolio, spanning infrastructure, development tools, and foundation models. Gartner appreciates the PAI-Lingjun platform’s capacity to support large-scale workloads on heterogeneous architectures. It also highlights Alibaba Cloud’s open-source tooling (EasyNLP, EasyCV, PAI-DSW, SONiC…) and the accompanying community support. It also notes the popularity of the open-weight Qwen models.

Beyond concerns around compliance and sovereignty that China’s nationality can raise, the interface and documentation can be more complex to master than those of American rivals. The geopolitical situation also deprives it of access to certain technologies, starting with accelerator chips.

AWS n’a pas l’architecture d’entraînement la plus efficace

AWS can rely on a mature cloud infrastructure whose reach, resilience and level of security/compliance Gartner highlights (notably the Nitro hypervisor and ISO 42001 certification). It is also a plus for the variety of managed infrastructure options, from Bedrock to SageMaker HyperPod.

On the flip side, this variety, with its levels of abstraction, control and responsibility, can complicate standardization at scale. AWS struggles more broadly to position its services within a clear and coherent narrative, according to Gartner. Deployment cycles can be longer as a result. The firm notes that network and training-cluster architectures are less efficient than those of other providers. The use of its Trainium chips may also require software optimization efforts.

Un risque de verrouillage chez Google…

Beyond its TPUs, Google stands out with the AI Hypercomputer architecture, which combines these chips with GPUs in an architecture featuring optimized networking and storage. Gartner also notes the available resources: by late 2025, Google was believed to hold around 25% of global AI compute capacity.

That said, access to this capacity—particularly NVIDIA GPUs—can still be difficult. Moreover, while Google has expanded support for frameworks such as Jax and PyTorch, optimization for TPUs may compromise workload portability. Also watch the pricing: a mix of token-based billing, infrastructure costs, deployment choices and commitment models can complicate cost attribution and forecasting.

… comme chez Huawei

Huawei benefits from controlling its stack (Ascend NPU, Kungpeng CPU, CANN architecture, MindSpore framework) and thus its cost‑performance ratio. It also has ModelArts, its development studio, which favors a use‑case-based approach built on sector-specific data to tailor Pangu models. Gartner appreciates how well it has integrated hybrid and private deployments into its business strategy.

As with Alibaba Cloud, Huawei’s Chinese nationality remains a hurdle for adoption in the Western world. As for its in-house stack, it may risk creating lock-in. Or at least imposing a higher learning curve for users accustomed to standards like CUDA. Caution is also warranted about the geographical footprint of its AI infrastructure, more limited than that of the other hyperscalers.

Microsoft: beware of cost governance

Gartner applauds Microsoft’s “unique value proposition” in GenAI through partnerships. It also notes the strong links between its AI infrastructure and the rest of Azure, as well as the exhaustiveness of its catalog of managed services. It cites Microsoft Foundry and Azure AI Landing Zone, which reduce the need for MLOps skills.

Like the other hyperscalers, Microsoft faces capacity-supply constraints. While it also has AI accelerators (Maia for inference), they currently play a limited role in its offering, still largely relying on third-party chips. The layers of services forming its catalog also complicate governance and cost forecasting.

Oracle’s tooling lacks maturity

Oracle stands out for the performance and elasticity of its OCI Supercluster architecture. It also scores points for deployment flexibility (Dedicated Region, Cloud@Customer, Oracle Database@…), and for its integration with other products (Oracle Fusion Cloud Applications, AI Data Platform, Autonomous Database…).

Oracle offers fewer integrations and community resources than its rivals; this may require more custom development. While the AI infrastructure is robust, the tooling to manage it (orchestration, observability…) is less mature and comprehensive. Documentation can be difficult to navigate, and support response times can vary.